Skip to contentSkip to conversation

Performance Comparison

DISCUSSION Aki M Laukkanen 1 participant 1 message May 1, 1999

lore
Share & tools
lore
#0 lore raw
Quoted 2 lines · Matthew Vanecek

Pardon my insolence, but perhaps someone could enlighten me as to what this has to do with compiling? From what I could read in the original

Alan probably misread the original message because HINT benchmark indeed doesn't test compiling speed of a compiler or anything like that. HINT is presented in this paper:

http://www.scl.ameslab.gov/Publications/HINT/ComputerPerformance.html

Basically the benchmark does a lot of heavy number crunching with a variable data size. As the figure in the original message showed, it is quite memory intensive.

As we were in compilers I ran the benchmark with two different versions of EGCS and different optimization flags. The results are shown by this figure:

http://www.cs.helsinki.fi/Aki.Laukkanen/hint-benchmark.gif

As with any other benchmark the compiler used does show in the results but not terribly much. To be completely certain the benchmark should be compiled with the same compiler, version and optimization flags which I think wasn't the case with Otto Solares' test runs but this can't explain all the differences.

The data size varies from about 160 bytes to over 40 megabytes with my configuration which is a P133 (classic)/256kB L2/32 MB ram running 2.2.5-ac6. The drop in the curve with very small data sizes I think is atleast partly resulting from the overhead caused by the gettimeofday() calls. This is a variable which varies from operating system to another.

The effect of the L1 cache can be seen quite clearly in this figure though. The performance starts dropping quite sharply at around 7500 bytes. This corresponds nicely with the cache size of 8kB. The missing 500 bytes can be well explained by stack usage etc. There is an odd quirk at 5500 bytes though, any ideas?

The curve for the segment covering the L2 cache is a lot less defined though. This is I think where page coloring could possibly help. Once it starts running from the main memory the curve is relatively stable until at about 25 MB when the curve drops sharply because of swapping.

Quoted 2 lines

light on, possible problems that could be addressed. Especially if the benchmarks were run with a credible benchmarks package.

Yes, I agree. I suggest more people to try the HINT benchmark. The code is ANSI-C and seems generally professional quite unlike the bytemark benchmark.

Signature · 6 lines
D.

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
Please read the FAQ at http://www.tux.org/lkml/